[ add Ralph ] On Tue, Feb 23, 2021 at 3:07 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Tue, Feb 23, 2021 at 02:48:20PM -0800, Dan Williams wrote: > > On Tue, Feb 23, 2021 at 10:54 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > > > > > On Tue, Feb 23, 2021 at 08:44:52AM -0800, Dan Williams wrote: > > > > > > > > The downside would be one extra lookup in dev_pagemap tree > > > > > for other pgmap->types (P2P, FSDAX, PRIVATE). But just one > > > > > per gup-fast() call. > > > > > > > > I'd guess a dev_pagemap lookup is faster than a get_user_pages slow > > > > path. It should be measurable that this change is at least as fast or > > > > faster than falling back to the slow path, but it would be good to > > > > measure. > > > > > > What is the dev_pagemap thing doing in gup fast anyhow? > > > > > > I've been wondering for a while.. > > > > It's there to synchronize against dax-device removal. The device will > > suspend removal awaiting all page references to be dropped, but > > gup-fast could be racing device removal. So gup-fast checks for > > pte_devmap() to grab a live reference to the device before assuming it > > can pin a page. > > From the perspective of CPU A it can't tell if CPU B is doing a HW > page table walk or a GUP fast when it invalidates a page table. The > design of gup-fast is supposed to be the same as the design of a HW > page table walk, and the tlb invalidate CPU A does when removing a > page from a page table is supposed to serialize against both a HW page > table walk and gup-fast. > > Given that the HW page table walker does not do dev_pagemap stuff, why > does gup-fast? gup-fast historically assumed that the 'struct page' and memory backing the page-table walk could not physically be removed from the system during its walk because those pages were allocated from the page allocator before being mapped into userspace. So there is an implied elevated reference on any page that gup-fast would be asked to walk, or pte_special() is there to "say wait, nevermind this isn't a page allocator page fallback to gup-slow()". pte_devmap() is there to say "wait, there is no implied elevated reference for this page, check and hold dev_pagemap alive until a page reference can be taken". So it splits the difference between pte_special() and typical page allocator pages. > Can you sketch the exact race this is protecting against? Thread1 mmaps /mnt/daxfile1 from a "mount -o dax" filesystem and issues direct I/O with that mapping as the target buffer, Thread2 does "echo "namespace0.0" > /sys/bus/nd/drivers/nd_pmem/unbind". Without the dev_pagemap check reference gup-fast could execute get_page(pte_page(pte)) on a page that doesn't even exist anymore because the driver unbind has already performed remove_pages(). Effectively the same percpu_ref that protects the pmem0 block device from new command submissions while the device is dying also prevents new dax page references being taken while the device is dying. This could be solved with the traditional gup-fast rules if the device driver could tell the filesystem to unmap all dax files and force them to re-fault through the gup-slow path to see that the device is now dying. I'll likely be working on that sooner rather than later given some of the expectations of the CXL persistent memory "dirty shutdown" detection.
